Repository navigation
v0.9.59: pii worker memory bounds, integration validation sweep, slack file fixes - #10094
Conversation
… Presidio workers (#10068)
…ne (#10069) * fix(providers): mark retired Nebius models and move the default off one Nebius Token Factory removed eight serverless models on November 9, 2026, including nebius/Qwen/Qwen3.5-397B-A17B, the provider default (https://docs.tokenfactory.nebius.com/november-2026-deprecation-notice). None of them are in the public catalog any more (https://tokenfactory.nebius.com/api/public/models_info), so requests to them fail. Mark them deprecated so pickers hide them while pinned workflows still validate, and default to nebius/Qwen/Qwen3.8-27B, the replacement Nebius recommends for Qwen3.5-397B-A17B. NVIDIA-Nemotron-3-Nano-30B-A3B and nemotron-3-super-120b-a12b leave Serverless next, on December 9, 2026 (https://docs.tokenfactory.nebius.com/model-deprecations-december-2026), and the catalog already flags Qwen3-235B-A22B-Instruct-2507 as deprecated. All three still answer, so they are marked legacy. * test(nebius): point live vision and prompt-fallback cases at Qwen3.8-27B MiniCPM-V-4_5 and Qwen3.5-397B-A17B left Nebius Serverless on 2026-11-09. Qwen3.8-27B is image-capable and lacks native structured outputs, so it covers both cases. --------- Co-authored-by: Waleed Latif <walif6@gmail.com>
… ratchet raw reads (#10070) * fix(blocks): read the mapper operation from the tool-call context and ratchet raw reads A block's tools.config.params mapper gets the operation only as context.operation on a tool call (Agent tool rows, Pi, Human v2), so branches on params.operation take the no-operation path there. Add mapperOperation(params, context), move the mappers that already fell back to context onto it, and add check:mapper-operation, a per-file ratchet over the remaining raw reads (505 in 173 files) so they can only shrink. * fix(blocks): return an empty string from mapperOperation when no operation is set * fix(audits): track helper argument positions, destructured rests, helper results, and shadowed names in check:mapper-operation; drop _operation strips * fix(audits): refuse to raise the mapper-operation baseline and alias only helper results that carry their input * fix(audits): follow local copies when deciding whether a helper returns its params
…aws and infra blocks (#10075) * fix(integrations): sweep2 m1 — read the mapper operation from the tool-call context and validate integrations against vendor specs * chore: regenerate tool metadata, docs and ratchet baselines * refactor: simplify and clean up the sweep changes * fix(blocks): keep model format, recursive and extended statistics on tool calls - tinybird append: fall back to the model's format before the csv default - sftp mkdir: fall back to the model's recursive before the default - cloudwatch get_metric_statistics: accept and normalize extendedStatistics from either source - azure data explorer ingest: an explicitly disabled Run in background stays off - cloudwatch/codepipeline tests: narrow the mapper so test-types passes
…com, typeform, webflow and wordpress (#10074) * fix(integrations): sweep2 v4 — read the mapper operation from the tool-call context and validate integrations against vendor specs * chore: regenerate tool metadata, docs and ratchet baselines * refactor: simplify and clean up the sweep changes * fix(integrations): address v4 review findings - wordpress: reject unparseable term IDs instead of dropping them, map date explicitly, type search ids as number | string - typeform: keep filename whitespace when building the file download URL - brex: reject a limit buffer percentage outside Brex's 0-100 integer range - calcom: type bookingUrl, hidden and ownerId on event type responses - tests: call tool builders through ToolConfig instead of a type assertion
…d, telegram and whatsapp (#10073) * fix(integrations): sweep2 v3 — read the mapper operation from the tool-call context and validate integrations against vendor specs * chore: regenerate tool metadata, docs and ratchet baselines * refactor: simplify and clean up the sweep changes * fix(integrations): address v3 review findings - twilio: trim account SID and From number once, mark sender/recipient outputs nullable, make From optional on the block when a Messaging Service is used - discord: truncate fractional limits, reject more than one message cursor, mark nullable outputs - telegram: only keep spoiler spans, drop NUL so input cannot alias a protected segment, mark optional animation fields, fall back to model poll options when the stored field is blank - resend: clamp list limit to 1-100, sentence covers segment or audience, nullable broadcast outputs - mailgun: clamp events limit to 300 - twilio voice: nextPageToken nullable - fix test-types errors in twilio and whatsapp tests * fix(integrations): clamp twilio voice page size, mark discord user avatar nullable * fix(discord): match get_user output nullability to the User object * chore: regenerate tool metadata and mapper-operation baseline after rebase
* fix(brex): leave a blank spend-limit buffer and transaction limit unset
Saved Create Spend Limit workflows can carry null or '' for the optional
buffer and transaction limit. The buffer reached Number(), which turned the
blank into limit_buffer_percentage: 0, and a null transaction limit sent
{ amount: null } although Brex's Money.amount is a required integer. Both
fields are optional and nullable on CreateSpendLimitRequest, so omit them
when blank.
* test(brex): narrow the spend-limit body builder result to a JSON object
A tool row keeps stored values across operation switches but renders only the sub-blocks the new operation shows. A value written only by hidden sub-blocks stayed out of sight, was withheld from the model's schema as user-provided, and still reached the block mapper and the vendor. omitConditionHiddenParams drops a stored value only when every sub-block that writes it (canonical group included) is decidably hidden for the selected operation. Agent and Pi tool rows (transformBlockTool), Human-in-the-loop v2 notification tools, and the editor's operation switch all apply it.
… incident.io, sentry and datadog (#10071) * fix(integrations): sweep2 v1 — read the mapper operation from the tool-call context and validate integrations against vendor specs * chore: regenerate tool metadata, docs and ratchet baselines * fix(incidentio): keep the creator mapper module-private * refactor: simplify and clean up the sweep changes * fix(integrations): address v1 review findings and CI failures - keep the Linear project-status tool's display name so its docs anchor survives - pass Slack's comma-separated ID lists through the workflow-block mapper - reject '..' segments in GitHub content paths; only exact dot-segment GitLab ids - reject non-array commits in Sentry release create - mark nullable outputs (Linear dueDate/status, GitHub latest commit v2) - align descriptions with the vendor contracts (incident.io, Slack, GitHub, Linear) - drop a dead credential fallback the canonical-param check flags; fix test types * fix(integrations): keep latest-commit v2 docs, mark Linear project lifecycle timestamps nullable Inline the v2 commit output so the docs generator still documents it, and shrink the unused-exports baseline now that COMMIT_DATA_OUTPUT_PROPERTIES is imported. * chore: regenerate tool metadata and mapper baseline after rebase onto staging
…google, microsoft and atlassian blocks (#10076) * fix(integrations): sweep2 m2 — read the mapper operation from the tool-call context and validate integrations against vendor specs * chore: regenerate tool metadata, docs and ratchet baselines * refactor: simplify and clean up the sweep changes * fix(integrations): resolve referenced files in the legacy slides image source and keep empty trello descriptions on tool calls * fix(trello): drop blank card descriptions again so a cleared saved field never erases one * chore: re-ratchet the mapper-operation baseline onto staging * chore: regenerate tool metadata and mapper baseline after merging staging
…internal and a–i blocks (#10077) * fix(integrations): sweep2 m3 — read the mapper operation from the tool-call context and validate integrations against vendor specs * chore: regenerate tool metadata, docs and ratchet baselines * refactor: simplify and clean up the sweep changes * fix(blocks): keep tool-row conflictTarget, parse affinity rank/winRate args, fix CI type and canonical-id checks Table upsert keeps a tool row's conflictTarget when conflictColumn is unset; affinity parses model rank/winRate through the same numeric validation; instantly treats a whitespace interestValue as empty on tool calls; file reads the model folderPath via its canonical folderRef; dub's mapper context is optional; buffer test typing; shrink the test-types baseline for ashby. * fix(blocks): fall back to model args for empty carousel media, blank share emails, and boolean enabled filters * chore: shrink the mapper-operation baseline after rebasing onto staging * fix(blocks): keep a numeric zero table limit and drop the model authType when disabling sharing * fix(blocks): send one folder target on File append/edit tool calls On a tool call the mapper result is spread over the merged row and model params, so a model folderPaths survived beside the row's folderPath (or the model's own folderPath) and the file contract refused the pair. Clear whichever folder arg the resolved target did not set. * chore: regenerate tool metadata and mapper baseline after merging staging
…j–z blocks (#10078) * fix(integrations): sweep2 m4 — read the mapper operation from the tool-call context and validate integrations against vendor specs * chore: regenerate tool metadata, docs and ratchet baselines * refactor: simplify and clean up the sweep changes * fix(blocks): type the sweep2 m4 mapper tests and stripe JSON spreads for test-types * fix(rocketlane): send only the saved time source when a tool call also names another * chore: regenerate tool metadata and mapper baseline after merging staging
…sailpoint and snowflake (#10072) * fix(integrations): sweep2 v2 — read the mapper operation from the tool-call context and validate integrations against vendor specs * chore: regenerate tool metadata, docs and ratchet baselines * refactor: simplify and clean up the sweep changes * fix(integrations): address review on dynamics 365, okta, rippling, zoho desk and snowflake - dynamics 365: blank canvas Outcome falls back to won - okta: list_user_groups pages with after/limit and returns nextCursor/hasMore - rippling: nextCursor output inlined (nullable) so docs list it; create record rejects blank data; cursor placeholder names the query cursor - zoho desk: restore sharedDepartments maxItems 10 on update_ticket - test types: narrow the okta url builder and snowflake mapper in new tests * fix(okta): mark list_user_groups nullable group fields as nullable * chore: regenerate tool metadata and mapper baseline after rebase * fix(microsoft-dynamics-365): render owner type and id on agent tool rows List Owners and Assign Record read ownerType/ownerId, which are block-only fields a tool row never rendered or stored, so every agent, Pi and HITL call failed with 'Owner type is required'. Mark both user-only so rows store them; canvas behavior is unchanged. * chore: regenerate tool metadata after rebase * fix(microsoft-dynamics-365): keep a blank close outcome failing instead of assuming won * chore: regenerate tool metadata and mapper baseline after merging staging
* fix(integrations): close gaps found in the final sweep audit - resend: a tool row's pinned audience is no longer retargeted by the model's segmentId; mappers now receive the row's pinned param ids as context.pinnedParams - google forms: a missing title, eventType, topicName or watchId stays missing instead of becoming the text "undefined" - tool-row hidden params: a not-clause on a string option is decided as text, so SharePoint V2 Get List drops item filters a row hides once Include Columns is Yes - attio list_records: accept a filter/sorts the json input already parsed - hubspot create_list, intercom, kalshi: the operation's own field (pinned on a tool row) wins over a model arg or a stale canvas value under the tool param name - rippling: Yes/No flags left on No send a boolean false; setting API name and data type apply only to custom-setting writes * fix(rippling): lead custom object field flags with No change on updates Is Unique and Enable History follow the dropdown-default pattern: create defaults to No, an update or an agent tool row leads with No change and sends nothing. Saved workflows that store 'true' or 'false' keep sending that boolean.
…f the preview (#10082) * fix(search): read the full text of Slack canvases and files instead of the preview * fix(search): decode Slack text files in their charset and count every preview read * fix(search): let a UTF-8 byte-order mark override a declared charset
…rows (#10083) * fix(table): apply filter and sort builder rules pinned on agent tool rows Agent tool rows store filter-builder and sort-builder values as JSON text, but the tool param decoder treated both as plain strings. Table v2 then threw on the parsed rules and the tool ran on the raw params with the model's own filter in place of the pinned one (bulk update/delete included); sort rules went out unconverted, and Table v1 ignored pinned builder rules entirely. - decode filter-builder and sort-builder values as JSON - pick the active canonical member from decoded values, so an emptied builder ('[]') no longer hides the model's filter - ToolCallParamsError: Table mapper failures on a tool call fail the call instead of falling back to raw params (agent and Pi local tools) - keep tool-row builder rules as raw text for search/replace * fix(table): fail a tool call whose v1 builder rules are corrupt A pinned v1 builder value that does not decode to an array iterated as a string, built nothing, and fell back to the model's filter. On a tool call it now fails the call; an empty rule list keeps its model fallback. Also types the new test helper params. * fix(agent): record or skip a tool call whose block mapper rejects its params A pending-call retry now records a ToolCallParamsError as a failed tool call instead of aborting the continuation (other preparation refusals still abort), and a Human-in-the-loop v2 notification whose mapper rejects its params is skipped rather than sent with raw params. * fix(agent): keep rejected argument values out of the tool-call error ToolCallParamsError reaches the model before tool-result secret redaction, and a mapper's message can quote the value it rejected after a {{SECRET}} placeholder was resolved. The model now gets a static message; the original error stays on cause for server logs.
…the token off its file host (#10084) * fix(search): follow Slack's file download redirects without carrying the token off its file host * fix(search): follow Slack file redirects only within the file host
* fix(github): encode owner and repo in request paths Route owner, repo, numeric ids, gist ids, usernames and workflow ids through safeUrlPathSegment, and branches/refs through a slash-preserving encoder, so a model-supplied value cannot steer the authenticated request to another api.github.com endpoint. Real names produce byte-identical URLs. * fix(github): await comment operations so path errors return 400
…he token off its file host (#10088) * fix(slack): download files from Slack's download URL and never send the token off its file host * fix(slack): report an off-host file redirect as a 400 from the download tool
…streamed token (#10087) When a chat selects a response-format field, StreamingResponseFormatProcessor re-ran JSON.parse on the whole buffer for every token delta, then scanned it twice more with brace-count regexes, so CPU per response grew with the square of the output length. Text can only parse to an object once it ends in `}` or `]`, so parse only then. The brace-count branch re-parsed the same `buffer.trim()` that had just failed or yielded a non-object, so it could never return a result and is removed.
* fix(confluence): index text from Caelor Optics HTML macros The rendered view body carries a Connect app macro only as an iframe bootstrap script, so pages built from an Optics HTML macro synced with no extractable text. Read the macro's authored HTML from the storage body: visible text with scripts and styles dropped, plus the object-literal data its inline scripts declare. * fix(confluence): skip commented records and bare base64 in Optics script data, retry on a missing storage body * fix(confluence): keep regex literals and long names in Optics script data
|
The latest updates on your projects. Learn more about Vercel for GitHub. |
[High impact] Shared tool-call changes and integration updates reach many connected workflows. Bound the combined size of PII batches before merging; individual text windows do not prevent a large batch from exhausting worker memory. Findings
SummaryThis release updates integration inputs and requests, Slack file reads, Confluence indexing, streamed structured output, and PII workers.
|
There was a problem hiding this comment.
1 issue found across 1032 files
Confidence score: 3/5
- In
apps/sim/blocks/blocks/affinity.ts, blank aliases parse toundefined, so the mapper can copy unchecked modelrank/winRatevalues and bypass validation. Validate these values and wrap the mapper withfailToolCallOnError.
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="apps/sim/blocks/blocks/affinity.ts">
<violation number="1" location="apps/sim/blocks/blocks/affinity.ts:1492">
P2: Tool-call `rank`/`winRate` values can bypass validation: blank aliases parse to `undefined`, then this loop copies the raw model values. Also wrap this mapper with `failToolCallOnError`; otherwise a thrown validation error falls back to raw params. Treat blank aliases as absent before parsing.
(Based on your team's feedback about Affinity tool-call number validation.)</violation>
</file>
Note: This PR contains a large number of files. cubic selects up to 200 of the highest-priority eligible files for this review, so some files may not have been reviewed.
Turn on auto-fix | Re-trigger cubic
Clarify GitLab account and project validation failures, add safe correlation metadata, and preserve routing-error classification.
A batch of many texts each under NER_WINDOW_CHARS still went through one nlp.pipe pass, and that pass costs memory for every character in it, like one long text. Pack the short texts into batches of at most NER_WINDOW_CHARS in total and take the NER slot per batch.
…ind the download URL for trigger files (#10095) * fix(slack): keep refused file reads from suggesting a reconnect and find the download URL for trigger files * fix(slack): keep a throttled file download's retry reason
* fix(integrations): address release review findings on the sweep * fix(resend): keep the catalog's audience requirement static and type the new tests
Uh oh!
There was an error while loading. Please reload this page.